Model Overview

This model, llmfan46/GLM-4-32B-0414-uncensored-heretic-v2, is a decensored version of the original zai-org/GLM-4-32B-0414 model, created by llmfan46 using the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method. The primary goal of this variant is to drastically reduce content refusals while preserving the original model's quality.

Key Differentiators

• Significantly Reduced Refusals: Achieves a 93% reduction in refusals (7/100) compared to the original model (100/100), making it highly uncensored.
• Quality Preservation: Maintains the original model's capabilities with a low KL divergence of 0.0555, indicating minimal degradation in performance.
• GLM-4-32B Base: Built upon the GLM-4-32B-0414 series, a 32 billion parameter model pre-trained on 15T high-quality data, including reasoning-type synthetic data.
• Enhanced Capabilities: The base GLM-4 model excels in instruction following, engineering code, artifact generation, function calling, search-based Q&A, and report generation.
• Function Calling Support: Demonstrates robust function calling capabilities, allowing integration with external tools in JSON format.

Ideal Use Cases

• Applications requiring minimal content restrictions or uncensored responses.
• Tasks involving complex code generation and engineering solutions.
• Scenarios demanding advanced function calling and tool integration.
• Search-based Q&A and detailed report generation.
• Use cases where preserving original model quality while removing censorship is critical.